Need help with Smooth_AP?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

Andrew-Brown1
136 Stars 17 Forks MIT License 30 Commits 5 Opened issues

Description

code for the ECCV '20 paper "Smooth-AP: Smoothing the Path Towards Large-Scale Image Retrieval"

Services available

!
?

Need anything else?

Contributors list

Smooth_AP

code for the ECCV '20 paper "Smooth-AP: Smoothing the Path Towards Large-Scale Image Retrieval"

The PyTorch implementation of the Smooth-AP loss function is found in src/SmoothAPloss.py

Training code and pre-trained weights coming soon...

teaser

Dependencies

  • Python 3.7.7
  • PyTorch 1.6.0
  • Cuda 10.1

Data

This repository is used for training using Smooth-AP loss on the following datasets:

  • PKU Vehicle ID (obtained from this website https://pkuml.org/resources/pku-vehicleid.html - must email authors for download permission)
  • INaturalist (2018 version - obtained from this website https://www.kaggle.com/c/inaturalist-2018/data)

We are the first to use the large-scale INaturalist dataset for the task of image retreival. The dataset splits can be downloaded here: https://drive.google.com/file/d/1sXfkBTFDrRU3_-NUs1qBP3sf0uMB98/view?usp=sharing . Unpack the zip into the INaturalist dataset directory.

Training the model

training results for the Vehicle ID and Inaturalist datasets can be replicated using this repository. To train the model on the Vehicle ID dataset, you can run:

  • python main.py --fclrmul 1 --bs 384

Paper

If you find this work useful, please consider citing:

@InProceedings{Brown20,
  author       = "Andrew Brown and Weidi Xie and Vicky Kalogeiton and Andrew Zisserman ",
  title        = "Smooth-AP: Smoothing the Path Towards Large-Scale Image Retrieval",
  booktitle    = "European Conference on Computer Vision (ECCV), 2020.",
  year         = "2020",
}

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.